Question about Run-time for Whisper

[amrix259]

Hello!

When I first started using Whisper it followed the time estimates that I see many people mentioning in various threads, about 1-3 minutes per 15 minutes of audio.

However, within the last day or two, this run-time has extended to multiple hours. I'm including the information in terminal as well as the file information. Anyone have any guesses on why this may have occurred or what I could do to fix it?

[jwnacnud]

Have you tried rebooting?

…

On Mon, Mar 6, 2023 at 1:04 PM amrix259 ***@***.***> wrote: Hello! When I first started using Whisper it followed the time estimates that I see many people mentioning in various threads, about 1-3 minutes per 15 minutes of audio. However, within the last day or two, this run-time has extended to multiple hours. I'm including the information in terminal as well as the file information. Anyone have any guesses on why this may have occurred or what I could do to fix it? [image: file_info_livingston_github] <https://user-images.githubusercontent.com/124815187/223218498-a4e392b4-9f89-4b13-8c6d-2c33a7f65cc5.png> [image: terminal_livingston_github] <https://user-images.githubusercontent.com/124815187/223218501-a3711317-3376-4bb1-a2bf-6e0c2a59d5f8.png> — Reply to this email directly, view it on GitHub <#1034>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGW5A5DYX2S55N3PCSDCR3W2Y7OHANCNFSM6AAAAAAVRSGAEU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[amrix259]

I rebooted and that resolved the issue :). Thank you!!

[jongwook]

Hi, the discrepancy you're seeing is likely the inference speed difference on GPU vs CPU. It appears that you're running whisper on CPU (the screenshot says FP16 is not supported on CPU; using FP32 instead), and the 1-3 minutes for 15-min audio is likely measured on a capable NVIDIA GPU.

You can install a CUDA-enabled pytorch if you have a Windows or Linux machine with an NVIDIA GPU, or use Google Colab which provides a GPU runtime. You can also venture with ROCm on AMD GPUs or MPS on ARM-based Macs, but these appear to be more error-prone.